Co-estimation of state of health and remaining useful life for lithium-ion batteries using a hybrid optimized framework

Abstract

Listen

Translate article

The accurate estimation of the state of health (SOH) and remaining useful life (RUL) in lithium-ion batteries (LIBs) is an important factor in assessing their operational characteristics such as robustness, efficiency, and effectiveness. The SOH and RUL are co-estimated to mitigate the LIB risk, limit unwanted fault occurrences, and ensure LIB safety. However, estimating the SOH and RUL is difficult considering the complex internal mechanism of the LIB and transients in operational performance due to aging. The currently available SOH and RUL estimation methods have demonstrated the inability to extract the right samples of data from battery parameters, choosing the right model hyperparameters and minimizing the capacity regeneration effect to produce reliable results. Furthermore, several research works using different optimization techniques suffer from premature convergence, local minima trap and slow convergence. Therefore, the novelty of this work is to deliver an improved hybrid optimized framework to co-estimate the SOH and RUL of LIBs which successfully addresses the above-mentioned limitations. The work utilizes a hybrid optimized gravitational search algorithm (GSA) and particle swarm optimization (PSO) model to optimize the cascaded forward neural network (CFNN) model parameters. First, the LIB dataset is acquired from the reliable NASA battery database, and its operational profiles are analysed to extract suitable parameters. Then, an appropriate data framework consisting of 31 samples is developed from the LIB using the systematic sampling method. The final dataset is developed considering four batteries which depicts high data dimensionality. The 31-dimensional data samples from the complete charging profile are acquired to train the proposed hybrid GSA-PSO-CFNN model for SOH estimation which attains improved search capability with fast convergence. Furthermore, the capacity regeneration effect from the outcomes with SOH estimation is minimized using the sliding window for RUL estimation. Moreover, the GSA-PSO-CFNN model is validated with other GSA-PSO optimized models and the MIT-Stanford battery dataset. High accuracy is achieved, with an average root mean square error (RMSE) and mean absolute percentage error (MAPE) for SOH and RUL estimation with the NASA dataset being 0.04715, 0.0225, and 2.3983, 1.4475, respectively, which confirms the high applicability of the GSA-PSO-CFNN model. The discussed co-estimation is applicable with the current battery management system (BMS) in electric vehicle (EV) applications for accurate estimation outcomes.